Mathematical Reasoning on AIME25 (ACC, mean@8)
70Accuracy (ACC)Qwen3-4B (Ours)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Qwen3-4B (Ours)Backbone=Qwen3-4B2026.04 | 70 | — | |
| Qwen3-8B (Ours)Backbone=Qwen3-8B2026.04 | 66.7 | — | |
| DS-Llama-8B (Ours)Backbone=DS-Llama-8B2026.04 | 43.3 | — | |
| DiPOBackbone=Qwen3-8B-Base2026.04 | 27.5 | 27.5 | |
| CDEBackbone=Qwen3-8B-Base2026.04 | 26.25 | 26.25 | |
| DAPOBackbone=Qwen3-8B-Base2026.04 | 25.83 | 25.83 | |
| DAPO w/ ELBackbone=Qwen3-8B-Base2026.04 | 25.42 | 25.42 | |
| DAPO w/ ELBackbone=Qwen3-4B-Base2026.04 | 24.58 | 24.58 | |
| DiPOBackbone=Qwen3-4B-Base2026.04 | 24.58 | 24.58 | |
| GRPOBackbone=Qwen3-8B-Base2026.04 | 24.58 | 24.58 | |
| CDEBackbone=Qwen3-4B-Base2026.04 | 24.17 | 24.17 | |
| DAPOBackbone=Qwen3-4B-Base2026.04 | 23.75 | 23.75 | |
| GRPOBackbone=Qwen3-4B-Base2026.04 | 23.33 | 23.33 | |
| All-Pairs+Backbone=qwen3-4b-base, nll coefficient=1.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.04 | 19.8 | — | |
| Margin+Backbone=qwen3-4b-base, nll coefficient=1.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.04 | 19.4 | — | |
| Margin+Backbone=qwen3-4b-base, nll coefficient=0.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.04 | 19 | — | |
| MPO+Backbone=qwen3-4b-base, nll coefficient=1.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.04 | 18.3 | — | |
| DPOBackbone=qwen3-4b-base, nll coefficient=1.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.04 | 18.1 | — | |
| Softmax+Backbone=qwen3-4b-base, nll coefficient=1.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.04 | 17.7 | — | |
| Softmax+Backbone=qwen3-4b-base, nll coefficient=0.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.04 | 16.9 | — | |
| DAPOBackbone=Qwen2.5-7B2026.04 | 16.67 | 16.67 | |
| DiPOBackbone=Qwen2.5-7B2026.04 | 16.67 | 16.67 | |
| MPO+Backbone=qwen3-4b-base, nll coefficient=0.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.04 | 16.2 | — | |
| All-Pairs+Backbone=qwen3-4b-base, nll coefficient=0.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.04 | 15.8 | — | |
| GRPOBackbone=Qwen2.5-7B2026.04 | 15.42 | 15.42 | |
| CDEBackbone=Qwen2.5-7B2026.04 | 15 | 15 | |
| DAPO w/ ELBackbone=Qwen2.5-7B2026.04 | 14.58 | 14.58 | |
| DPOBackbone=qwen3-4b-base, nll coefficient=0.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.04 | 14.2 | — | |
| RFTBackbone=qwen3-4b-base, nll coefficient=0.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.04 | 12.9 | — | |
| Base modelBackbone=Qwen3-8B-Base2026.04 | 9.17 | 9.17 | |
| Base modelBackbone=Qwen3-4B-Base2026.04 | 3.75 | 3.75 | |
| Base modelBackbone=Qwen2.5-7B2026.04 | 2.08 | 2.08 |